SAS Data Analysis Examples
Zero-Inflated Poisson Regression

Version info: Code for this page was tested in SAS 9.3

Zero-inflated poisson regression is used to model count data that has an excess of zero counts. Further, theory suggests that the excess zeros are generated by a separate process from the count values and that the excess zeros can be modeled independently.  Thus, the zip model has two parts, a poisson count model and the logit model for predicting excess zeros. You may want to review these Data Analysis Example pages, Poisson Regression and Logit Regression.

Please Note: The purpose of this page is to show how to use various data analysis commands. It does not cover all aspects of the research process which researchers are expected to do. In particular, it does not cover data cleaning and verification, verification of assumptions, model diagnostics and potential follow-up analyses.

Examples of zero-inflated Poisson regression

Example 1.  School administrators study the attendance behavior of high school juniors over one semester at two schools.  Attendance is measured by number of days of absent and is predicted by gender of the student and standardized test scores in math and language arts.  Many students have no absences during the semester.

Example 2. The state wildlife biologists want to model how many fish are being caught by fishermen at a state park. Visitors are asked whether or not they have a camper, how many people were in the group, were there children in the group and how many fish were caught. Some visitors do not fish, but there is no data on whether a person fished or not. Some visitors who did fish did not catch any fish so there are excess zeros in the data because of the people that did not fish.

Description of the data

Let's pursue Example 2 from above using the dataset fish.sas7bdat

We have data on 250 groups that went to a park.  Each group was questioned about how many fish they caught (count), how many children were in the group (child), how many people were in the group (persons), and whether or not they brought a camper to the park (camper).   

In addition to predicting the number of fish caught, there is interest in predicting the existence of excess zeros, i.e., the zeroes that were not simply a result of bad luck fishing. We will use the variables child, persons, and camper in our model. Let's look at the data.

proc means data = fish mean std min max var;
  var count child persons;

The MEANS Procedure

Variable            Mean         Std Dev         Minimum         Maximum        Variance
count          3.2960000      11.6350281               0     149.0000000     135.3738795
child          0.6840000       0.8503153               0       3.0000000       0.7230361
persons        2.5280000       1.1127303       1.0000000       4.0000000       1.2381687

proc univariate data = fish noprint;
  histogram count / midpoints = 0 to 50 by 1 vscale = count ;

proc freq data = fish;
  tables camper;

The FREQ Procedure

                                   Cumulative    Cumulative
camper    Frequency     Percent     Frequency      Percent
     0         103       41.20           103        41.20
     1         147       58.80           250       100.00


Analysis methods you might consider

Below is a list of some analysis methods you may have encountered. Some of the methods listed are quite reasonable while others have either fallen out of favor or have limitations.

SAS zero-inflated Poisson regression analysis using proc genmod

If you are using SAS version 9.2 or higher, you can run a zero-inflated Poisson model using proc genmod.

The output begins with a summary of the model and the data.  This is followed by a list of goodness of fit statistics. 

The next block of output includes parameter estimates from the count portion of the model.  It also includes the standard errors, Wald 95% confidence intervals, Wald Chi-square statistics, and p-values for the parameter estimates. 

The last block of output corresponds to the zero-inflation portion of the model.  This is a logistic model predicting the zeroes.  The output includes parameter estimates for the inflation model predictors and their standard errors, Wald 95% confidence intervals, Wald Chi-square statistics, and p-values. 

All of the predictors in both the count and inflation portions of the model are statistically significant. This model fits the data significantly better than the null model, i.e., the intercept-only model. To show that this is the case, we can run the null model (a model without any predictors) and compare the null model with the current model using chi-squared test on the difference of log likelihoods.

proc genmod data = fish;
  model count =  /dist=zip;
  zeromodel  / link = logit ;
                                       The GENMOD Procedure

                                        Model Information

             Data Set                         WORK.FISH    Written by SAS

             Distribution          Zero Inflated Poisson
             Link Function                           Log
             Dependent Variable                    count

                             Number of Observations Read         250
                             Number of Observations Used         250

                              Criteria For Assessing Goodness Of Fit

                 Criterion                     DF           Value        Value/DF

                 Deviance                               2254.0459
                 Scaled Deviance                        2254.0459
                 Pearson Chi-Square           248       1918.7890          7.7371
                 Scaled Pearson X2            248       1918.7890          7.7371
                 Log Likelihood                          679.4854
                 Full Log Likelihood                   -1127.0229
                 AIC (smaller is better)                2258.0459
                 AICC (smaller is better)               2258.0945
                 BIC (smaller is better)                2265.0888

           Algorithm converged.

                       Analysis Of Maximum Likelihood Parameter Estimates

                                  Standard     Wald 95% Confidence          Wald
   Parameter    DF    Estimate       Error           Limits           Chi-Square    Pr > ChiSq

   Intercept     1      2.0316      0.0349      1.9631      2.1000       3388.16        <.0001
   Scale         0      1.0000      0.0000      1.0000      1.0000

NOTE: The scale parameter was held fixed.

                Analysis Of Maximum Likelihood Zero Inflation Parameter Estimates

                                  Standard     Wald 95% Confidence          Wald
   Parameter    DF    Estimate       Error           Limits           Chi-Square    Pr > ChiSq

   Intercept     1      0.2728      0.1277      0.0225      0.5232          4.56        0.0327

The log likelihoods for the full model and null mode are -1031.6084 and -1127.0229, respectively. The chi-squared value is 2*( -1031.6084 - -1127.0229) = 190.829. Since we have three predictor variables in the full model, the degrees of freedom for the chi-squared test is 3. This yields a p-value <.0001.  Thus, our overall model is statistically significant.

We may want to compare the current zero-inflated poisson model with the plain poisson model, which can be done with the Vuong test. Currently, the Vuong test is not a standard part of proc genmod, but a macro program that performs the Vuong test is available from SAS here.  Usage of the macro program requires the %include statement, in which we list the location of the macro. This macro program takes quite a few arguments, as shown below. We rerun the models to get produce these required input arguments. With the zero-inflated poisson model, there are total of 3 regression parameters and with the plain poisson regression model a total of two regression parameters. The scale parameter is the dispersion parameter from each corresponding model, and for our poisson model is fixed at 1.

%inc "C:\Users\ALIN\Documents\My SAS Files\";
proc genmod data = fish order=data;
  class camper;
  model count = child camper /dist=zip;
  zeromodel persons;
  output out=outzip pred=predzip pzero=p0;
  store m1;
proc genmod data = outzip order=data;
  class camper;
  model count = child camper /dist=poi;
  output out=out pred=predpoi;
%vuong(data=out, response=count,
       model1=zip, p1=predzip, dist1=zip, scale1=1.00, pzero1=p0, 
       model2=poi, p2=predpoi, dist2=poi, scale2=1.00,
       nparm1=3,   nparm2=2)

                                         The Vuong Macro
	     			 	Model Information

                           Data Set                         out
                           Response                         count
                           Number of Observations Used      250
                           Model 1                          zip
                              Distribution                  ZIP
                              Predicted Variable            predzip
                              Number of Parameters          3
                              Scale                         1.00
                              Zero-inflation Probability    p0
                              Log Likelihood                -1031.6084
                           Model 2                          poi
                              Distribution                  POI
                              Predicted Variable            predpoi
                              Number of Parameters          2
                              Scale                         1.00
                              Log Likelihood                -1358.5929
                                            Vuong Test
                          H0: models are equally close to the true model
                        Ha: one of the models is closer to the true model

                       Vuong Statistic            Z    Pr>|Z|      Model

                       Unadjusted            3.5814    0.0003       zip
                       Akaike Adjusted       3.5705    0.0004       zip
                       Schwarz Adjusted      3.5512    0.0004       zip
                                         Clarke Sign Test
                          H0: models are equally close to the true model
                        Ha: one of the models is closer to the true model

                       Clarke Statistic           M    Pr>=|M|      Model

                       Unadjusted           13.0000    0.1137        zip
                       Akaike Adjusted       2.0000    0.8496        zip
                       Schwarz Adjusted      2.0000    0.8496        zip

For the Vuong test, a significant z indicates that the zero-inflated model is better. Here we see that the preferred model is a zero-inflated poisson model over a regular poisson model.

We can use the estimate statement to help understand our model. We will compute the expected counts for the categorical variable camper while holding the continuous variable child at its mean value using the atmeans option, as well as calculate the predicted probability that an observation came from the zero-generating process.  In the estimate statement, we provide values at which to evaluate each coefficient for both the poisson model and the zero-inflation model.  The sets of coefficients of the two models are separated by the @ZERO keyword.

proc genmod data = fish;
  class camper;
  model count = child camper /dist=zip;
  zeromodel persons /link = logit ;
  estimate "camper = 0" intercept 1 child .684 camper 1 0 @ZERO intercept 1 persons 2.528; 
  estimate "camper = 1" intercept 1 child .684 camper 0 1 @ZERO intercept 1 persons 2.528; 

                                    Contrast Estimate Results

                                   Mean            Mean              L'Beta    Standard
Label                          Estimate      Confidence Limits     Estimate       Error     Alpha

camper = 0                       2.4220      1.9724      2.9741      0.8846      0.1048      0.05
camper = 0 (Zero Inflation)      0.4677      0.3838      0.5536     -0.1292      0.1756      0.05
camper = 1                       5.5768      4.8823      6.3701      1.7186      0.0679      0.05
camper = 1 (Zero Inflation)      0.4677      0.3838      0.5536     -0.1292      0.1756      0.05

                                    Contrast Estimate Results

                                                 L'Beta             Chi-
           Label                            Confidence Limits     Square    Pr > ChiSq

           camper = 0                       0.6792      1.0899     71.28        <.0001
           camper = 0 (Zero Inflation)     -0.4735      0.2150      0.54        0.4619
           camper = 1                       1.5856      1.8516    641.42        <.0001
           camper = 1 (Zero Inflation)     -0.4735      0.2150      0.54        0.4619

In the Mean Estimate column, we find predicted counts of fish from the poisson model, ignoring the zero-inflation model, for both camper = 0 and camper = 1, as well as the predicted probability of belonging to the zero-generating process from the zero-inflation model.  The zero-inflation model does not include camper as a predictor, so the probability of zero for both zero-inflation models is the same.  To get the expected counts of fish from the mixture of the two models, simply multiply the expected counts from the poisson model by the probability of getting a non-zero from the zero-inflation model (1 - p(zero)).  Thus, the expected counts of fish for camper = 0 and camper = 1 including zero-inflation are 2.422*(1-.04677) = 1.289 and 5.5768*(1-.04677) = 2.968, respectively.

SAS zero-inflated Poisson analysis using proc countreg

Proc countreg is another option for running a zero-inflated Poisson regression in SAS (again, version 9.2 or higher).  This procedure allows for a few more options specific to count outcomes than proc genmod.  The proc countreg code for the original model run on this page appears below.  We indicate method = qn to specify the quasi-Newton optimization process that matches the proc genmod results. 

proc countreg data = fish method = qn;
  class camper;
  model count = child camper / dist= zip;
  zeromodel count ~ persons;

                                      The COUNTREG Procedure

                                     Class Level Information

                                  Class       Levels    Values

                                  camper           2    0 1

                                        Model Fit Summary

                            Dependent Variable                  count
                            Number of Observations                250
                            Data Set                       MYLIB.FISH
                            Model                                 ZIP
                            ZI Link Function                 Logistic
                            Log Likelihood                      -1032
                            Maximum Absolute Gradient      3.69075E-7
                            Number of Iterations                   13
                            Optimization Method          Quasi-Newton
                            AIC                                  2075
                            SBC                                  2096

                 Algorithm converged.

                                       Parameter Estimates

                                                       Standard                 Approx
            Parameter        DF        Estimate           Error    t Value    Pr > |t|

            Intercept         1        2.431911        0.041271      58.93      <.0001
            child             1       -1.042838        0.099988     -10.43      <.0001
            camper 0          1       -0.834022        0.093627      -8.91      <.0001
            camper 1          0               0               .        .         .
            Inf_Intercept     1        1.297439        0.373850       3.47      0.0005
            Inf_persons       1       -0.564347        0.162962      -3.46      0.0005

SAS Zero-inflated Poisson analysis using proc nlmixed

For those using a version of SAS prior to 9.2, a zero-inflated negative binomial model is doable, though significantly more difficult.  Please see this code fragment: Zero-inflated Poisson and Negative Binomial Using Proc Nlmixed.

Things to consider

See also



How to cite this page

Report an error on this page or leave a comment

The content of this web site should not be construed as an endorsement of any particular web site, book, or software product by the University of California.